Gartner expects that by 2023, one-third of companies that have implemented IoT will also have implemented AI in conjunction with at least one IoT project. The popularity and intrigue in AI show no signs of slowing down, with it being named one of the 5 biggest tech trends this year. We support our customers by integrating AI into their products, which in turn, pushes their products to new technological heights and generates invaluable data to improve future work.
Thaumatec can help you be at the forefront of the AI revolution. We can support your company throughout the entire process: from the initial conception to solution development and production.
Chief Technology Officer
Contact me to talk about AI!
OUR AI KNOWLEDGE
Computer vision: object detection, semantic segmentation, image generation; techniques: various architectures of CNN, GAN, transfer learning, autoencoders, TensorFlow, TensorFlow Lite;
• Natural Language Processing: speech recognition, NL understanding like text summarization, topic modelling or sentiment analysis; techniques: TFiDF, Word2Vec, BERT, GPT-3 and many more.
• Predictive modeling: time series forecasting, classification, regression; techniques: ARIMA, regressions, random forests, Xgboost, deep learning, and many more.
• Optimization: Genetic Algorithms, Bayesian Optimization;
• Recommendation engines: collaborative, content based, hybrids.
• Anomalies detection: clustering, dimensionality reduction, isolation forests.
• Simulations: Monte-Carlo, reinforcement learning.
• Software development: Python (Pandas, NumPy, Scikit-learn), R;
• Data Visualization (Matplotlib, Bokeh, Tableau, d3).
• All kind of databases (sql, nosql), data warehouses (cloud, on-premises), data lakes and data transformation tools;
• Cloud IoT tool stacks: Azure IoT Hub, AWS IoT Core;
• Big data tool stack: Hadoop, Kafka, HDInsights, Spark, Dask;
• Software development in general (Python).
• ML models training and operationalisation: Azure Machine Learning Studio, Amazon SageMaker;
• Devops tooling: CI/CD tools, Docker, Kubernetes;
• Software development in general (Python).
AI Consulting – we can help you explore the “art of the possible”
Data Engineering – extracting, transforming, cleaning, joining and loading the data from various sources
Machine Learning and Statistical Modelling – computer vision, forecasting, recommendation engines, anomalies detection, reinforcement learning, optimization, simulations
Model operationalization – deployment and monitoring in the cloud and on the edge
Data visualization – the fastest path to learning from the data is the proper visualization
OUR AI FRAMEWORK
The main goal of this phase is to achieve a common understating between the customer and the data science team. The first step is to help the team to understand the business domain. Next, the detailed definition of what goals the model is to achieve needs to be agreed – the use case is to be specified. With this, the team needs to review the available data sets – quantity and quality. Finally, the scope for the Proof of Data is chosen, with expected accuracy.
2. Proof of Data
The main goal of this phase is the use case feasibility. With Data Science, usually the most important aspect to verify is weather the data is good enough to produce an accurate enough model. The approach is to choose a small part of the problem, initially clean the data and apply a small subset of standard models. The result is next analysed and the decision about the path of continuation is agreed with the customer.
The delivery phase of a data science project is highly iterative. (It is based on the common standard CRISP-DM). Understanding of the business domain and data sets of the modelled aspect allows to start the data cleaning (preparation). The prepared data allow to train initial models. Looking at the models accuracy allows the team to understand potential issues in the data and results in more data cleaning. After multiple iterations, the results are evaluated with the customer. This usually provides to better understanding of the business and data set, allowing for creation of the next iteration of models. This cycle repeats until the accuracy achieve the acceptable level and can be deployed (operationalised) into a working solution.
Once deployed, the model needs to be monitored. Often, the accuracy of the model degrades with time, which is the result of the introduction of the new types of input data (not known when the model was trained) and the shifts of the business process that is modelled. Additionally, the working model could benefit from feedback loops or other data that can be started to gather after deployment. A common practise is to re-train the models recurrently (or when the accuracy drops below a given threshold) and to re-design them occasionally.
CNN, TensorFlow, TFiDF, Word2Vex, XgBoost, ARIMA, Clustering, Isolation forests, Python, Scikit learn, R
Azure IoT Hub, AWS IoT Core, SQL, noSQL, Azure ML Studio, Hadoop, Kafka, Docker, Kubernetes